Multi-Scale Image Processing via Variable Kernel Filtering

The patent segments the image processing task into multiple scale levels by applying filtering kernels of different sizes (e.g., 3x3, 5x5, 7x7). Each kernel size processes specific feature scales, allowing the system to handle different image details separately and reconstruct a comprehensive enhanced image that maintains both fine and coarse structures while suppressing noise effectively.

The patent introduces the dimension of kernel size variation to traditional filtering. Instead of using a single fixed kernel size, the system processes the image with multiple kernel sizes simultaneously, adding a new degree of freedom (kernel size parameter) to the filtering operation. This enables the system to capture features at different scales and reconstruct images with improved quality and noise suppression.

The patent divides the image processing into separate filtering operations for different scale ranges. Small kernels (e.g., 3x3) handle fine details and high-frequency features, medium kernels (e.g., 5x5) handle intermediate structures, and large kernels (e.g., 7x7) handle coarse low-frequency components. This segmentation allows each kernel to be optimized for its specific scale range, improving adaptability while managing complexity through structured organization.

The patent systematically varies the kernel size parameter across multiple filtering operations. By changing this single parameter (kernel size) while keeping the filtering mechanism consistent, the system achieves multi-scale feature enhancement without fundamentally changing the processing architecture. This parameter-based approach maintains relatively simple processing logic while achieving versatile multi-scale处理能力.

The patent segments the noise suppression task across multiple kernel sizes, where small kernels preserve fine details and large kernels suppress broader noise patterns. By distributing the noise suppression function across different scale levels rather than relying on a single kernel, the system achieves more effective noise removal while maintaining processing efficiency through parallel or sequential application of optimized kernels for each scale.

The patent applies filtering operations that may appear excessive (multiple kernels on the same image) but each kernel performs a specific partial function. Small kernels handle fine-detail noise, medium kernels handle intermediate noise, and large kernels handle coarse noise. This partial action approach ensures that no single kernel is overloaded, maintaining processing speed while collectively achieving superior noise suppression through the combined effect of multiple specialized filters.